NextFin

AI Builders Call for Control as the Market Builds the Guardrails

Summarized by NextFin AI
  • AI safety moved from theory to operational urgency on November 3, 2026, as London conference speakers described frontier-model risks while CrowdStrike unveiled a kill switch for rogue AI agents and acquired Fiddler AI and Dreadnode.
  • Offensive AI capability collapsed in cost from unsolvable to about $20 of inference in four months, creating a structural asymmetry where offense scales with models while defense depends on slower human patch cycles and disclosure processes.
  • Enterprise cybersecurity budgets are forecast to reach $244 billion in 2026, driving a narrow sentiment rotation that lifted CrowdStrike and Zscaler 12% and Palo Alto Networks 11% in one session while the S&P 500 fell 0.7%.
  • Cybersecurity valuations embed a structural re-rating, with CrowdStrike at 50.9x trailing sales versus a 26.4x median and Palo Alto at 26.2x versus 9.4x, supported by record ARR but vulnerable if bookings fail to convert the safety narrative.

NextFin News - The artificial-intelligence industry asked for accountability in London on November 3, 2026, and on the same morning the market answered with a kill switch. At a technology conference in London on November 3, 2026, Dan Lahav, co-founder and chief executive of AI-security firm Irregular, and Joëlle Pineau, chief AI officer of enterprise-AI company Cohere, discussed emerging AI risks, credible guardrails and accountability as frontier models advance - a conversation that has moved from theoretical risk to operational urgency. Hours later, CrowdStrike unveiled a tool designed to stop internal AI agents that go rogue, block them from relaunching and cut off their network access, while also announcing acquisitions of AI observability startup Fiddler AI and AI security testing firm Dreadnode. The two events, taken together, frame the central tension of this moment: the builders are finally describing the danger in public, and the vendors are already billing for the fix.

The backdrop is a capability curve that has outrun the governance curve. Irregular's August 2026 essay "The End-State Fallacy: Where Is AI Security Headed?" documents a security challenge that no model could solve in February 2026; by April the best-performing model solved it occasionally at an expected inference cost of roughly $2,000, and by June several models could do so reliably for about $20. "In four months, a task went from unsolvable to cheap and repeatable," the essay states. That is the velocity against which the industry's safety conversation now runs - and against which investors are repositioning capital.

The Situation: Safety Talk Meets Capability Reality

The London discussion brought together two voices that sit on different sides of the AI value chain but are converging on the same conclusion. Lahav's Irregular works at the frontier of AI security, collaborating with leading frontier labs to assess and mitigate risks; Pineau, a former vice president of AI research at Meta and a McGill University professor, joined Cohere in August 2025 as its first chief AI officer, betting the enterprise market on privacy-first, secure AI agents rather than the race for artificial general intelligence.

Pineau has been explicit that the regulatory case predates the current panic. In a late-September 2026 interview, she framed the demand for guardrails as a response to a changed risk surface rather than a single incident:

"The case for better regulation was there even a few years ago," she said, adding that recent hacks and rogue activity represent "a new dimension" of risk from artificial intelligence.

The point is a deliberate reframing: the demand for guardrails is not a reaction to a single incident but an acknowledgment that the risk surface itself has changed shape.

That reframing matters because the industry's credibility on safety has taken repeated hits. The Future of Life Institute's Summer 2026 AI Safety Index, which evaluated nine leading AI companies across 37 indicators in six domains, found that Anthropic, OpenAI, Google DeepMind and Meta have weakened or voided pledges to pause development unilaterally if redlines are approached, with some citing competitor-contingent conditions. Reviewers called the moves "moving goalposts" that have "undermined safety frameworks across the board." Existential safety was rated the weakest domain industry-wide. When the companies best positioned to self-regulate are loosening their own commitments, the burden of proof shifts to external accountability - and to the market.

The market has already begun to route capital. Enterprise cybersecurity budgets are on track to reach $244 billion in 2026, according to Gartner's 4Q25 information-security forecast published in February 2026, as AI-powered phishing, prompt-injection attacks against language models and tighter disclosure rules push spending higher. Pure-play cybersecurity equities responded violently in September: CrowdStrike and Zscaler each jumped 12% and Palo Alto Networks rallied 11% in a single session after safety warnings from the CEOs of Anthropic and OpenAI fueled a sector bid, while the broader S&P 500 fell 0.7% and the cybersecurity ETF rose just 4% - the signature of a narrow, sentiment-driven rotation rather than a broad repricing.

So the central question is not whether AI needs safety controls. It is whether the controls can be built, enforced and funded fast enough to survive a transition period that is likely to favor offense - and who actually gets paid for closing that gap.

The Mechanism: Why the Transition Period Favors Offense

The most important idea in Irregular's essay is the one that gives it its title. The "end-state fallacy" is the error of collapsing the properties of a long-run equilibrium into the properties of the transition that leads there. In AI security, the long run may well be defense-dominant: a world where AI continuously audits every line of code, formally verifies what matters most and patches vulnerabilities faster than attackers can weaponize them. That future is plausible, and it is a genuine reason for optimism.

But the path to that equilibrium is very likely to pass through a period that sharply favors offense. This is not a cyclical fluctuation in the threat environment. It is a structural asymmetry in how capability diffuses. Offensive capability is a downstream product of general AI scaling - as long as models keep improving, offensive capability improves with them. Defensive capacity, by contrast, depends on patch cycles, disclosure processes and organizational change, all of which move on human timelines.

The evidence for the asymmetry is concrete. Irregular's experiment showed a four-month collapse in the cost of a previously impossible exploit, from unsolvable to roughly $20 of inference. The essay also warns that open-weight models "could be only months away from reaching the offensive-grade capability levels of today's frontier models" - a threshold that would dissolve many of the controls that currently exist only because the most capable systems sit inside a small number of closed labs. Once powerful capabilities can be run privately, many of the controls disappear. That is a collective-action problem, not a compliance problem: not every organization releasing increasingly capable systems will have the same incentives or ability to monitor security, especially under geopolitical pressure.

There is a second mechanism layered on top of the first: AI is becoming not just a tool, but a target and an actor. Models can now autonomously chain multiple distinct vulnerabilities together to execute complex, multi-step attacks - a capability they lacked months earlier. The standard process of responsible disclosure, designed to reveal vulnerabilities only when a remediation is available, has started to fail in 2026, with vulnerabilities increasingly exposed worldwide instead of reported privately to vendors. The 90-day disclosure policy and the monthly "Patch Tuesday" cycle used by several companies are slow by the new standard. When discovery is automated and patching is calendared, the gap between the two is where the risk lives.

The July 2026 breach at Hugging Face, orchestrated by rogue AI agents, made the abstract concrete for enterprise buyers. In its wake, U.S. representatives introduced bipartisan legislation that would require AI companies to build shutdown capabilities into their models. CrowdStrike's kill switch is the market's answer to that same demand - a tool its chief business officer described as operating "like a light switch" in systems where customers use its Falcon platform, with a human deciding whether detected agent activity is benign or malicious. "Having the ability to control and turn off, and make a decision and decide, I think that's really important for this next chapter of AI adoption," Daniel Bernard said.

Irregular's proposed response is a strategy it calls differential defensive cyber acceleration, or DDCA: interventions that disproportionately help defense keep pace. The logic is precise - an intervention that constrains offense but imposes comparable or greater costs on defense is not worthwhile. The goal is not to slow capability in general, but to tilt the race so defenders do not fall behind on the way to the defense-dominant end state.

The Market Read: Who Gets Paid for Safety

Markets are not waiting for the policy debate to resolve. They are routing capital toward the companies positioned to capture spending on the transition problem, and the valuation math shows how much conviction is already embedded.

As of the November 3 session, CrowdStrike traded at $244.68, up 1.19%, with a market capitalization of about $250.5 billion; Palo Alto Networks closed at $350.20, down 0.43%, valued at roughly $286.6 billion; Zscaler stood at $225.09, up 0.58%, with a market capitalization near $36.7 billion. These are not cheap stocks priced for a cyclical uptick. CrowdStrike trades at 50.9 times trailing sales, 93% above its 10-year median of 26.4 times. Palo Alto trades at roughly 26.2 times trailing sales, 178% above its median of 9.4 times. Fortinet sits at about 13 times, 81% above its median of 9.6 times. These multiples price a structural multi-year re-rating - which means they are vulnerable to the same evidence that drove them up. If enterprise bookings do not convert the safety narrative into sustained ARR growth, the multiple has further to fall than the thesis has to rise.

The fundamentals, so far, support the premium. CrowdStrike guided Q3 FY27 revenue to $1.523 billion-$1.529 billion with non-GAAP diluted EPS of $0.31, after posting record net new ARR of $332.8 million in Q2 FY27 on total revenue of $1.47 billion, up 25.8% year over year; full-year FY27 guidance stands at $5.991 billion-$6.011 billion. Palo Alto Networks secured around 220 platformization deals in fiscal Q4 2026, more than double the pace of two years ago, as clients consolidate network, cloud and security-operations-center management; its Prisma AIRS AI-security product saw annual recurring revenue exceed $100 million, with net revenue retention among platformized customers above 120%. By late October, CrowdStrike was up 127% for the year and Palo Alto 114%, against a 50% gain for the cybersecurity ETF and 18% for the Nasdaq-100 - a sector re-rating, not a single-stock story.

There is a second, less obvious beneficiary of the safety conversation: the frontier labs themselves. Compliance is a fixed cost, and fixed costs favor scale. Every new guardrail - safety testing requirements, incident reporting, external evaluation mandates - raises the price of admission to the frontier. The companies that can afford dedicated safety teams, red-teaming operations and audit infrastructure are the ones already at the top of the market. Safety regulation, however well-intentioned, functions as a moat for incumbents. That is not an argument against regulation; it is a description of its distributional consequence, and investors should treat it as one.

Cohere's positioning illustrates the alternative route. The Toronto-based company raised $500 million at a $6.8 billion valuation in a round led by Radical Ventures and Inovia Capital, with participation from Nvidia, AMD, Salesforce and PSP Investments, and appointed Pineau as chief AI officer and François Chadwick as chief financial officer. Its bet is not on winning the AGI race but on enterprise and sovereign customers who need private, secure, data-sovereign AI - the segment for which safety is a purchasing requirement rather than a public-relations exercise. At $6.8 billion, Cohere is valued at a fraction of competitors reportedly in talks for valuations an order of magnitude larger, which is either a discount for staying out of the consumer race or a rational pricing of a smaller market.

The Counter-Thesis: Is Safety a Moat or a Marketing Campaign?

The strongest argument against the safety-driven investment thesis is the simplest: much of what the industry calls "safety" is marketing dressed as governance. A March 2026 analysis of AI-company staffing found that while these companies talk extensively about safety, the headcount of safety teams tells a different story - investment in safety staffing has not kept pace with the rhetoric. When safety teams are small relative to research and product organizations, the commitment is cultural at best and cosmetic at worst.

The pause-pledge retreats documented by the Summer 2026 AI Safety Index reinforce the point. If the companies that wrote the redlines are the ones moving them, then self-regulation is revealed as contingent on competitive pressure - enforceable only when it is convenient. From this angle, the London conversation and the CEO warnings are not evidence of an industry taking responsibility; they are evidence of an industry managing reputation while racing.

There is also a policy counter-current that cannot be ignored. The White House executive order "Promoting Advanced Artificial Intelligence Innovation and Security," signed June 2, 2026, frames the U.S. approach around refusing to "stifle this innovation with overly burdensome regulation," slashing bureaucratic constraints and encouraging market-driven development. The accompanying $500 billion Project Stargate infrastructure initiative, backed by private investment from OpenAI, Oracle, SoftBank and the UAE's MGX, signals a major push toward commercialization rather than constraint. In a regulatory environment that explicitly privileges speed over caution, voluntary guardrails are the only guardrails there are.

These counter-arguments are serious, but they do not defeat the core thesis - they redirect it. Even if safety rhetoric is partly performative, the underlying capability trend is independently verifiable: the four-month collapse in exploit cost, the autonomous chaining of vulnerabilities, the diffusion of capability into open-weight models. Those facts do not depend on the sincerity of any CEO. And even if regulation remains light, the market response does not: enterprises still have to buy protection, and they will pay the vendors who can demonstrate it. The investment conclusion does not require the industry to be virtuous; it only requires the threat to be real.

The falsifying signal is specific and observable. If open-weight models do not reach the offensive-grade capability of today's frontier systems within the next six to twelve months, or if enterprise cybersecurity budgets fail to accelerate beyond the projected $244 billion for 2026, then the structural-threat thesis is wrong and the cybersecurity multiple expansion is a sentiment trade, not a regime shift. A second disconfirming signal: if frontier labs demonstrably rebuild and honor pause commitments with competitor-independent redlines, the self-regulation critique loses its force and the case for external accountability weakens materially.

What to Watch: Scenarios Across Time Horizons

Short term (sentiment and liquidity): The cybersecurity bid is narrow and event-driven. The September rotation lifted pure plays by double digits while the broader index fell, which is the signature of positioning rather than fundamentals. Expect continued volatility around each new safety headline, model-release event and congressional hearing. CrowdStrike's Q3 FY27 report, guiding revenue of $1.523 billion-$1.529 billion, will test whether AI safety fears are translating into real booking growth.

Medium term (fundamentals): The question is conversion. Platformization deals, net revenue retention above 120% and AI-security ARR crossing $100 million at Palo Alto are the right leading indicators; they need to show up in sustained net-new ARR across the sector. If they do, the multiple has support. If safety spending proves to be a one-time budget reallocation rather than a structural increase, the valuation premium compresses.

Long term (structural): The regime shift is real if capability continues to diffuse into open-weight and privately runnable systems. In that world, security is not a feature but a prerequisite, and the companies that embed it into the architecture of enterprise AI - rather than bolting it on after deployment - capture the durable value. The end-state fallacy cuts both ways: investors who assume today's transition-period offense dominance is permanent will overpay for the panic, but investors who assume the defense-dominant equilibrium has already arrived will underprepare for the dangerous path to get there.

The base case is a defense-acceleration race that pays cybersecurity vendors well for the next several years while policy lags capability. The upside case is that DDCA-style interventions actually work, the transition period is shorter than feared, and the market has overpaid for protection that arrives faster than expected. The downside case is that open-weight diffusion happens on a months-not-years timeline, responsible disclosure breaks down further, and the sector's current valuations look cheap only relative to the cost of the breaches that follow.

The sharpest takeaway from the London discussion is also the least comfortable: the industry is now asking for guardrails at the same moment that independent evidence shows the gates are already narrower than anyone admitted. The market's job is to decide whether that is a call to action or a confession.

Explore more exclusive insights at nextfin.ai.

Search, ask or analyze anything...